# Guptasiddharth

> The personal tech blog and portfolio of Siddharth Gupta, a software engineer and tech entrepreneur. The writing focuses on building real AI systems (agent harnesses, agentic payments, AI safety, on-prem coding agents, and inference infrastructure) alongside hard-won lessons in scaling high-throughput engineering backends, trading systems, and blockchain products to millions of users.

## Key pages

- [Home](https://www.guptasiddharth.com/): Siddharth Gupta on engineering, AI agents, and systems.
- [Posts](https://www.guptasiddharth.com/posts/): The full archive of technical writing.
- [Tags](https://www.guptasiddharth.com/tags/): Posts grouped by topic.
- [About](https://www.guptasiddharth.com/about/): Background, expertise, and ways to get in touch.

## AI systems series

- [How I Built an AI Harness That Ships Products on Its Own](https://www.guptasiddharth.com/posts/how-i-built-an-ai-harness/): A multi-agent orchestration harness that takes a one-line task and returns a deployed product for about five cents. Specialized agents, computed iteration budgets, guardrails, and a learning file that turns each failure into a permanent rule.
- [A Coding Agent That Proves Its Fixes and Never Leaves Your Network](https://www.guptasiddharth.com/posts/sovereign-coding-agent/): A verification-first, air-gapped coding agent that runs on the customer's own hardware, is model-agnostic, and never claims a fix works until the repo's own tests pass. The harness, not the model, is the moat.
- [The Inference Substrate, or Why Every AI App Pays a Tax It Cannot See](https://www.guptasiddharth.com/posts/the-inference-substrate/): A thesis on the missing layer between applications and AI models that cuts cost and latency through automated distillation, speculative execution, and outcome-learned routing, and why model companies structurally cannot build it.
- [Proof of Mandate, or How to Hijack an Agent and Steal Nothing](https://www.guptasiddharth.com/posts/proof-of-mandate/): A cryptographically enforced safety model that makes a compromised agent's malicious action impossible to sign rather than merely unlikely, with a demo where a fully jailbroken agent tries to drain a treasury and fails on-chain.
- [A Paywall for AI Agents, Built on a Forgotten HTTP Code](https://www.guptasiddharth.com/posts/a-paywall-for-ai-agents-http-402/): Using HTTP 402 to let AI agents pay for resources on their own, with on-chain settlement, non-extractable keys, a policy engine, and stateless verification that runs on an edge function.
- [Separating Signal from Noise on YouTube with One Claude Call](https://www.guptasiddharth.com/posts/separating-signal-from-noise-on-youtube/): A tool that reads a video transcript, hands it to Claude once, and returns only the substantive claims, data, and arguments, filtering out sponsor reads and hype.

## Engineering and systems

- [How I scaled a trading engine to 100K TPS](https://www.guptasiddharth.com/posts/scaling-a-crypto-matching-engine/): Migrating a cryptocurrency exchange's matching engine to Go for high concurrency, low latency, and optimal CPU utilization.
- [Optimizing MySQL Indexing for Maximum Performance](https://www.guptasiddharth.com/posts/optimizing-mysql-indexing-for-maximum-performance/): A practical guide to MySQL indexing, the pitfalls of over-indexing, and choosing the right columns to accelerate query retrieval.